What not to paste into AI
Real leak cases, what each AI vendor keeps, and how to redact before you share.
What a redaction tool cannot see in your text
Eight rule groups and 28 patterns can reliably find supported API key formats but cannot identify a confidential project name. The distinction is not a gap in the rules; it is the boundary of the whole method.
Your .env file hides its secrets from most redaction tools
DB_PASSWORD and PASSWORD look the same to a person and completely different to a regular expression. We shipped this bug ourselves, and here is the one character that caused it.
The fifteen-second check before you hit send
Automated rules handle the formats. This is the part only you can do, in the order that catches the most for the least effort.
We ran 100,000 random card numbers through the Luhn check. 10% passed.
Card detection is the one rule PromptScrub ships switched off. Here is the measurement behind that decision, and what it means for the order numbers in your support transcripts.
The AI tool your company has not approved is the one you are using
Shadow AI is not a discipline problem. It is what happens when the sanctioned option is slower than the unsanctioned one, and the fix is not another policy email.
Why the detection runs in your browser, and how to verify it
A redaction tool that uploads your text to check it has reproduced the problem it claims to solve. Here is the architecture, and the thirty-second check that proves it.
Summarising support tickets with AI, without shipping the customer
Transcripts are the densest personal data most teams handle, and the most tempting thing to automate. What redaction handles well here, and the two things it will not touch.
Asking AI to debug your config without handing over the keys
A worked example: fifteen lines of deploy failure, eight secrets found, and the redacted version that still gets you a useful answer.
What actually happens to the text you paste into an AI chat
Not a scare story. A plain reading of what the major providers say they do with your input, why the defaults differ between consumer and API access, and which distinction actually matters.
A private key is the one thing you cannot half-redact
PEM blocks are the easiest secret to detect and the worst to leak. Our rule takes all 163 characters of a test key, and here is why partial masking is disabled for this one type.
Which redaction style keeps the AI useful?
The same sentence, three ways. Placeholders preserve the structure the model reasons over; removal destroys it. Here is the output of each, side by side.
Your hostnames are a map of your network
db01.internal tells a reader you have at least one more database server. We tested eleven host formats to see which ones leave the building.